Using HPWREN camera images for fire ignition detection

June 1, 2020

HPWREN staff have been discussing fire plume detection via its camera infrastructure with various projects for a number of years. Over the last few years technologies in the areas of Deep Learning and Edge Computing have matured to a point where real-time image analysis at the origin location of the images is becoming feasible. An example HPWREN experiment is described at http://hpwren.ucsd.edu/news/20190823/. The following article is by Wei Shung Chung, a collaborator who is separately organizing multiple volunteer competitions (hackathons) to use artificial intelligence technologies that have promising potential for our societal needs, including wildfire detection.

Note: This is about work in progress, with no operational service at this time. Constructive comments, suggestions and interest in participation can be submitted via the "Contact and Feedback Form" on the main HPWREN web page, and will be forwarded to the article author.

Contributed by Wei Shung Chung, Founder of AI For Mankind

AI For Mankind https://aiformankind.org/ is a 501(c)(3) nonprofit organization with the mission of mobilizing the tech community to work on world challenging problems using AI and Data. We organize tech talks, workshops, and hackathons. We want to build a grassroot community of volunteers creating solutions using AI and Data to bring positive impacts to society at large.

Smoke Detection Model

AI For Mankind has built a smoke detection model using bounding box annotated HPWREN camera images. In this train dataset, we annotated 744 images with bounding boxes out of total 3292 images. We trained a Faster R-CNN with Resnet-101 wildfire smoke detector using the above annotated HPWREN images.

The results look promising but it can be improved further. Here are some examples:

We tested the model against this fire 20200306_FIRE_mlo-n-mobo-c
The detection model did picked up the smoke at 1583516059_+00180

Here are the results. This basic object detection model shows promising results. We can definitely make it even better. Hope the hackathon participants will help improve the model or come up with a better model.

Missed out 1583515939_+00060: False Negative

Missed out 1583515999_+00120: False Negative

Detected smoke at 1583516059_+00180. See the tiny green bounding box above: True Positive

Also detected smoke in 1583516119_+00240: True Positive

Also detected smoke in 1583516119_+00240: True Positive

Another Detected Smoke: True Positive
20200304_FIRE_rm-w-mobo-c
1583346786_+01740 first_detected

However, we also noticed some false positives as shown below.

AI For Mankind launched the first Lets Stop Wildfires Hackathon https://aiformankind.org/lets-stop-wildfires-hackathon in 2019. The theme of the first hackathon was mainly focused on classifying HPWREN image as smoke vs non-smoke. Quickstart end to end colab notebooks were provided to the participants.

For this 2020 AI For Mankind’s Lets Stop Wildfires Hackathon 2.0 https://aiformankind.org/lets-stop-wildfires-hackathon-2.0, we formulated the detection problem as object detection (classifying and localizing) and segmentation problems. We also engage our volunteers to help annotate more HPWREN images.

We also provide the quick start codebase on how to build the same smoke detection model we had built above.